Introduction
The inspection of reinforced concrete (RC) structures is critical for ensuring public safety and infrastructure longevity. Traditionally, this process has been manual, labor-intensive, and subjective. While Convolutional Neural Networks (CNNs) have revolutionized automated damage detection, they face a significant hurdle: "catastrophic forgetting." When a model is trained on a new task (e.g., identifying corrosion), it often loses its ability to perform previously learned tasks (e.g., crack detection).
To combat this, engineers typically resort to "Joint Training," which requires retraining the model on all cumulative data, or maintaining multiple specialized models. Both approaches are resource-heavy, demanding massive storage and increasing prediction latency. This paper introduces the Continual Learning-based Damage Recognition Model (CLDRM), a framework designed to accumulate knowledge across multiple structural health monitoring tasks without the need for original raw data from previous training cycles.
Methodology: The CLDRM Framework
The CLDRM integrates a ResNet-34 backbone with Learning without Forgetting (LwF) techniques. The primary mechanism for retaining knowledge is Knowledge Distillation. In this setup, the model acts as both a "teacher" (the version containing old knowledge) and a "student" (the version learning new tasks).
Instead of relying on old images, the student model is trained using the teacher’s "soft targets"—the probability distributions of its outputs. A specialized loss function ensures the student's predictions for old tasks do not stray significantly from the teacher's baseline. Additionally, the researchers implemented a "warm-up" phase, where only the parameters for the new task are updated initially. This prevents the initial random weights of a new task from corrupting the established feature extractors of the network.
Dataset and Classification Tasks
The framework was validated using a comprehensive dataset of 16,840 images (224 × 224 pixels), including data from the PEER Hub ImageNet. The images represent a wide variety of structural conditions and components.
The learning process was divided into four distinct recognition tasks:
- Spalling Check: A binary classification (Spalling vs. No Spalling).
- Damage Level: Categorizing damage as Undamaged, Minor, or Heavy.
- Component Type: Identifying the structural element (Wall, Column, or Beam).
- Damage Type: Distinguishing between specific failure modes like Shear cracks, Flexural cracks, Alkali-Silica Reaction (ASR), and Corrosion.
Key Insights and Results
The study yielded several critical insights into the behavior of continual learners in structural health monitoring:
- Optimal Distillation Temperature: A distillation temperature ($T$) of 5 was found to be most effective. This parameter smooths the probability distribution, allowing the model to capture subtle visual cues in low-probability outputs that are essential for distinguishing between similar damage types.
- Feature Fusion and Task Order: The researchers observed that learning related tasks (e.g., spalling and ASR) can be synergistic. The sequence of learning matters; the model performs best when the most complex tasks are introduced later, allowing them to benefit from the rich feature set developed during simpler tasks.
- Efficiency Gains: Compared to using four separate models, the CLDRM reduced data storage requirements and prediction time by approximately 75%.
- Performance Metrics: The model achieved an average final accuracy of 89.32%. While slightly lower than the "Joint Training" gold standard, it significantly outperformed standard fine-tuning. Specifically, while fine-tuning saw an 8% accuracy drop on the first task after four training rounds, CLDRM limited this "forgetting" to just 5%.
Conclusion and Future Directions
The CLDRM represents a significant step toward "lifelong learning" for structural inspection AI. By reducing the computational and storage overhead, it paves the way for real-time deployment on edge devices like UAVs (drones) for automated bridge or building inspections.
Future research will focus on improving the detection of "Minor Damage," which currently suffers from visual ambiguity with undamaged surfaces. The authors also propose integrating Generative Adversarial Networks (GANs) for "generative replay" and exploring Long Short-Term Memory (LSTM) layers to manage knowledge drift over even longer sequences of tasks.
Summary updated